<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Overview
🤗 Diffusers provides a collection of training scripts for you to train your own diffusion models. You can find all of our training scripts in [diffusers/examples](https://github.com/huggingface/diffusers/tree/main/examples).
Each training script is:
-**Self-contained**: the training script does not depend on any local files, and all packages required to run the script are installed from the `requirements.txt` file.
-**Easy-to-tweak**: the training scripts are an example of how to train a diffusion model for a specific task and won't work out-of-the-box for every training scenario. You'll likely need to adapt the training script for your specific use-case. To help you with that, we've fully exposed the data preprocessing code and the training loop so you can modify it for your own use.
-**Beginner-friendly**: the training scripts are designed to be beginner-friendly and easy to understand, rather than including the latest state-of-the-art methods to get the best and most competitive results. Any training methods we consider too complex are purposefully left out.
-**Single-purpose**: each training script is expressly designed for only one task to keep it readable and understandable.
Our current collection of training scripts include:
| Training | SDXL-support | LoRA-support | Flax-support |
|---|---|---|---|
| [unconditional image generation](https://github.com/huggingface/diffusers/tree/main/examples/unconditional_image_generation)[](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) | | | |
These examples are **actively** maintained, so please feel free to open an issue if they aren't working as expected. If you feel like another training example should be included, you're more than welcome to start a [Feature Request](https://github.com/huggingface/diffusers/issues/new?assignees=&labels=&template=feature_request.md&title=) to discuss your feature idea with us and whether it meets our criteria of being self-contained, easy-to-tweak, beginner-friendly, and single-purpose.
## Install
Make sure you can successfully run the latest versions of the example scripts by installing the library from source in a new virtual environment:
Then navigate to the folder of the training script (for example, [DreamBooth](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth)) and install the `requirements.txt` file. Some training scripts have a specific requirement file for SDXL, LoRA or Flax. If you're using one of these scripts, make sure you install its corresponding requirements file.
```bash
cd examples/dreambooth
pip install-r requirements.txt
# to train SDXL with DreamBooth
pip install-r requirements_sdxl.txt
```
To speedup training and reduce memory-usage, we recommend:
- using PyTorch 2.0 or higher to automatically use [scaled dot product attention](../optimization/torch2.0#scaled-dot-product-attention) during training (you don't need to make any changes to the training code)
- installing [xFormers](../optimization/xformers) to enable memory-efficient attention
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Stable Diffusion XL
<Tipwarning={true}>
This script is experimental, and it's easy to overfit and run into issues like catastrophic forgetting. Try exploring different hyperparameters to get the best results on your dataset.
</Tip>
[Stable Diffusion XL (SDXL)](https://hf.co/papers/2307.01952) is a larger and more powerful iteration of the Stable Diffusion model, capable of producing higher resolution images.
SDXL's UNet is 3x larger and the model adds a second text encoder to the architecture. Depending on the hardware available to you, this can be very computationally intensive and it may not run on a consumer GPU like a Tesla T4. To help fit this larger model into memory and to speedup training, try enabling `gradient_checkpointing`, `mixed_precision`, and `gradient_accumulation_steps`. You can reduce your memory-usage even more by enabling memory-efficient attention with [xFormers](../optimization/xformers) and using [bitsandbytes'](https://github.com/TimDettmers/bitsandbytes) 8-bit optimizer.
This guide will explore the [train_text_to_image_sdxl.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py) training script to help you become more familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Then navigate to the example folder containing the training script and install the required dependencies for the script you're using:
```bash
cd examples/text_to_image
pip install-r requirements_sdxl.txt
```
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell, like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
## Script parameters
<Tip>
The following sections highlight parts of the training script that are important for understanding how to modify it, but it doesn't cover every aspect of the script in detail. If you're interested in learning more, feel free to read through the [script](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_sdxl.py) and let us know if you have any questions or concerns.
</Tip>
The training script provides many parameters to help you customize your training run. All of the parameters and their descriptions are found in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L129) function. This function provides default values for each parameter, such as the training batch size and learning rate, but you can also set your own values in the training command if you'd like.
For example, to speedup training with mixed precision using the bf16 format, add the `--mixed_precision` parameter to the training command:
```bash
accelerate launch train_text_to_image_sdxl.py \
--mixed_precision="bf16"
```
Most of the parameters are identical to the parameters in the [Text-to-image](text2image#script-parameters) training guide, so you'll focus on the parameters that are relevant to training SDXL in this guide.
-`--pretrained_vae_model_name_or_path`: path to a pretrained VAE; the SDXL VAE is known to suffer from numerical instability, so this parameter allows you to specify a better [VAE](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)
-`--proportion_empty_prompts`: the proportion of image prompts to replace with empty strings
-`--timestep_bias_strategy`: where (earlier vs. later) in the timestep to apply a bias, which can encourage the model to either learn low or high frequency details
-`--timestep_bias_multiplier`: the weight of the bias to apply to the timestep
-`--timestep_bias_begin`: the timestep to begin applying the bias
-`--timestep_bias_end`: the timestep to end applying the bias
-`--timestep_bias_portion`: the proportion of timesteps to apply the bias to
### Min-SNR weighting
The [Min-SNR](https://huggingface.co/papers/2303.09556) weighting strategy can help with training by rebalancing the loss to achieve faster convergence. The training script supports predicting either `epsilon` (noise) or `v_prediction`, but Min-SNR is compatible with both prediction types. This weighting strategy is only supported by PyTorch and is unavailable in the Flax training script.
Add the `--snr_gamma` parameter and set it to the recommended value of 5.0:
```bash
accelerate launch train_text_to_image_sdxl.py \
--snr_gamma=5.0
```
## Training script
The training script is also similar to the [Text-to-image](text2image#training-script) training guide, but it's been modified to support SDXL training. This guide will focus on the code that is unique to the SDXL training script.
It starts by creating functions to [tokenize the prompts](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L478) to calculate the prompt embeddings, and to compute the image embeddings with the [VAE](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L519). Next, you'll a function to [generate the timesteps weights](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L531) depending on the number of timesteps and the timestep bias strategy to apply.
Within the [`main()`](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L572) function, in addition to loading a tokenizer, the script loads a second tokenizer and text encoder because the SDXL architecture uses two of each:
The [prompt and image embeddings](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L857) are computed first and kept in memory, which isn't typically an issue for a smaller dataset, but for larger datasets it can lead to memory problems. If this is the case, you should save the pre-computed embeddings to disk separately and load them into memory during the training process (see this [PR](https://github.com/huggingface/diffusers/pull/4505) for more discussion about this topic).
After calculating the embeddings, the text encoder, VAE, and tokenizer are deleted to free up some memory:
```py
deltext_encoders,tokenizers,vae
gc.collect()
torch.cuda.empty_cache()
```
Finally, the [training loop](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/text_to_image/train_text_to_image_sdxl.py#L943) takes care of the rest. If you chose to apply a timestep bias strategy, you'll see the timestep weights are calculated and added as noise:
If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Once you’ve made all your changes or you’re okay with the default configuration, you’re ready to launch the training script! 🚀
Let’s train on the [Naruto BLIP captions](https://huggingface.co/datasets/lambdalabs/naruto-blip-captions) dataset to generate your own Naruto characters. Set the environment variables `MODEL_NAME` and `DATASET_NAME` to the model and the dataset (either from the Hub or a local path). You should also specify a VAE other than the SDXL VAE (either from the Hub or a local path) with `VAE_NAME` to avoid numerical instabilities.
<Tip>
To monitor training progress with Weights & Biases, add the `--report_to=wandb` parameter to the training command. You’ll also need to add the `--validation_prompt` and `--validation_epochs` to the training command to keep track of results. This can be really useful for debugging the model and viewing intermediate results.
[PyTorch XLA](https://pytorch.org/xla) allows you to run PyTorch on XLA devices such as TPUs, which can be faster. The initial warmup step takes longer because the model needs to be compiled and optimized. However, subsequent calls to the pipeline on an input **with the same length** as the original prompt are much faster because it can reuse the optimized graph.
print(f'Inference time is {time()-start} sec after compilation')
```
</hfoption>
</hfoptions>
## Next steps
Congratulations on training a SDXL model! To learn more about how to use your new model, the following guides may be helpful:
- Read the [Stable Diffusion XL](../using-diffusers/sdxl) guide to learn how to use it for a variety of different tasks (text-to-image, image-to-image, inpainting), how to use it's refiner model, and the different types of micro-conditionings.
- Check out the [DreamBooth](dreambooth) and [LoRA](lora) training guides to learn how to train a personalized SDXL model with just a few example images. These two training techniques can even be combined!
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# T2I-Adapter
[T2I-Adapter](https://hf.co/papers/2302.08453) is a lightweight adapter model that provides an additional conditioning input image (line art, canny, sketch, depth, pose) to better control image generation. It is similar to a ControlNet, but it is a lot smaller (~77M parameters and ~300MB file size) because its only inserts weights into the UNet instead of copying and training it.
The T2I-Adapter is only available for training with the Stable Diffusion XL (SDXL) model.
This guide will explore the [train_t2i_adapter_sdxl.py](https://github.com/huggingface/diffusers/blob/main/examples/t2i_adapter/train_t2i_adapter_sdxl.py) training script to help you become familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Then navigate to the example folder containing the training script and install the required dependencies for the script you're using:
```bash
cd examples/t2i_adapter
pip install-r requirements.txt
```
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell, like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
<Tip>
The following sections highlight parts of the training script that are important for understanding how to modify it, but it doesn't cover every aspect of the script in detail. If you're interested in learning more, feel free to read through the [script](https://github.com/huggingface/diffusers/blob/main/examples/t2i_adapter/train_t2i_adapter_sdxl.py) and let us know if you have any questions or concerns.
</Tip>
## Script parameters
The training script provides many parameters to help you customize your training run. All of the parameters and their descriptions are found in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L233) function. It provides default values for each parameter, such as the training batch size and learning rate, but you can also set your own values in the training command if you'd like.
For example, to activate gradient accumulation, add the `--gradient_accumulation_steps` parameter to the training command:
```bash
accelerate launch train_t2i_adapter_sdxl.py \
----gradient_accumulation_steps=4
```
Many of the basic and important parameters are described in the [Text-to-image](text2image#script-parameters) training guide, so this guide just focuses on the relevant T2I-Adapter parameters:
-`--pretrained_vae_model_name_or_path`: path to a pretrained VAE; the SDXL VAE is known to suffer from numerical instability, so this parameter allows you to specify a better [VAE](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)
-`--crops_coords_top_left_h` and `--crops_coords_top_left_w`: height and width coordinates to include in SDXL's crop coordinate embeddings
-`--conditioning_image_column`: the column of the conditioning images in the dataset
-`--proportion_empty_prompts`: the proportion of image prompts to replace with empty strings
## Training script
As with the script parameters, a walkthrough of the training script is provided in the [Text-to-image](text2image#training-script) training guide. Instead, this guide takes a look at the T2I-Adapter relevant parts of the script.
The training script begins by preparing the dataset. This incudes [tokenizing](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L674) the prompt and [applying transforms](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L714) to the images and conditioning images.
Within the [`main()`](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L770) function, the T2I-Adapter is either loaded from a pretrained adapter or it is randomly initialized:
The [optimizer](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L952) is initialized for the T2I-Adapter parameters:
```py
params_to_optimize=t2iadapter.parameters()
optimizer=optimizer_class(
params_to_optimize,
lr=args.learning_rate,
betas=(args.adam_beta1,args.adam_beta2),
weight_decay=args.adam_weight_decay,
eps=args.adam_epsilon,
)
```
Lastly, in the [training loop](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L1086), the adapter conditioning image and the text embeddings are passed to the UNet to predict the noise residual:
If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Now you’re ready to launch the training script! 🚀
For this example training, you'll use the [fusing/fill50k](https://huggingface.co/datasets/fusing/fill50k) dataset. You can also create and use your own dataset if you want (see the [Create a dataset for training](https://moon-ci-docs.huggingface.co/docs/diffusers/pr_5512/en/training/create_dataset) guide).
Set the environment variable `MODEL_DIR` to a model id on the Hub or a path to a local model and `OUTPUT_DIR` to where you want to save the model.
Download the following images to condition your training with:
To monitor training progress with Weights & Biases, add the `--report_to=wandb` parameter to the training command. You'll also need to add the `--validation_image`, `--validation_prompt`, and `--validation_steps` to the training command to keep track of results. This can be really useful for debugging the model and viewing intermediate results.
prompt="pale golden rod circle with old lace background"
generator=torch.manual_seed(0)
image=pipeline(
prompt,image=control_image,generator=generator
).images[0]
image.save("./output.png")
```
## Next steps
Congratulations on training a T2I-Adapter model! 🎉 To learn more:
- Read the [Efficient Controllable Generation for SDXL with T2I-Adapters](https://huggingface.co/blog/t2i-sdxl-adapters) blog post to learn more details about the experimental results from the T2I-Adapter team.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Text-to-image
<Tipwarning={true}>
The text-to-image script is experimental, and it's easy to overfit and run into issues like catastrophic forgetting. Try exploring different hyperparameters to get the best results on your dataset.
</Tip>
Text-to-image models like Stable Diffusion are conditioned to generate images given a text prompt.
Training a model can be taxing on your hardware, but if you enable `gradient_checkpointing` and `mixed_precision`, it is possible to train a model on a single 24GB GPU. If you're training with larger batch sizes or want to train faster, it's better to use GPUs with more than 30GB of memory. You can reduce your memory footprint by enabling memory-efficient attention with [xFormers](../optimization/xformers). JAX/Flax training is also supported for efficient training on TPUs and GPUs, but it doesn't support gradient checkpointing, gradient accumulation or xFormers. A GPU with at least 30GB of memory or a TPU v3 is recommended for training with Flax.
This guide will explore the [train_text_to_image.py](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) training script to help you become familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Then navigate to the example folder containing the training script and install the required dependencies for the script you're using:
<hfoptionsid="installation">
<hfoptionid="PyTorch">
```bash
cd examples/text_to_image
pip install-r requirements.txt
```
</hfoption>
<hfoptionid="Flax">
```bash
cd examples/text_to_image
pip install-r requirements_flax.txt
```
</hfoption>
</hfoptions>
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell, like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
## Script parameters
<Tip>
The following sections highlight parts of the training script that are important for understanding how to modify it, but it doesn't cover every aspect of the script in detail. If you're interested in learning more, feel free to read through the [script](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image.py) and let us know if you have any questions or concerns.
</Tip>
The training script provides many parameters to help you customize your training run. All of the parameters and their descriptions are found in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L193) function. This function provides default values for each parameter, such as the training batch size and learning rate, but you can also set your own values in the training command if you'd like.
For example, to speedup training with mixed precision using the fp16 format, add the `--mixed_precision` parameter to the training command:
```bash
accelerate launch train_text_to_image.py \
--mixed_precision="fp16"
```
Some basic and important parameters include:
-`--pretrained_model_name_or_path`: the name of the model on the Hub or a local path to the pretrained model
-`--dataset_name`: the name of the dataset on the Hub or a local path to the dataset to train on
-`--image_column`: the name of the image column in the dataset to train on
-`--caption_column`: the name of the text column in the dataset to train on
-`--output_dir`: where to save the trained model
-`--push_to_hub`: whether to push the trained model to the Hub
-`--checkpointing_steps`: frequency of saving a checkpoint as the model trains; this is useful if for some reason training is interrupted, you can continue training from that checkpoint by adding `--resume_from_checkpoint` to your training command
### Min-SNR weighting
The [Min-SNR](https://huggingface.co/papers/2303.09556) weighting strategy can help with training by rebalancing the loss to achieve faster convergence. The training script supports predicting `epsilon` (noise) or `v_prediction`, but Min-SNR is compatible with both prediction types. This weighting strategy is only supported by PyTorch and is unavailable in the Flax training script.
Add the `--snr_gamma` parameter and set it to the recommended value of 5.0:
```bash
accelerate launch train_text_to_image.py \
--snr_gamma=5.0
```
You can compare the loss surfaces for different `snr_gamma` values in this [Weights and Biases](https://wandb.ai/sayakpaul/text2image-finetune-minsnr) report. For smaller datasets, the effects of Min-SNR may not be as obvious compared to larger datasets.
## Training script
The dataset preprocessing code and training loop are found in the [`main()`](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L490) function. If you need to adapt the training script, this is where you'll need to make your changes.
The `train_text_to_image` script starts by [loading a scheduler](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L543) and tokenizer. You can choose to use a different scheduler here if you want:
Then the script [loads the UNet](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L619) model:
Next, the text and image columns of the dataset need to be preprocessed. The [`tokenize_captions`](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L724) function handles tokenizing the inputs, and the [`train_transforms`](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L742) function specifies the type of transforms to apply to the image. Both of these functions are bundled into `preprocess_train`:
Lastly, the [training loop](https://github.com/huggingface/diffusers/blob/8959c5b9dec1c94d6ba482c94a58d2215c5fd026/examples/text_to_image/train_text_to_image.py#L878) handles everything else. It encodes images into latent space, adds noise to the latents, computes the text embeddings to condition on, updates the model parameters, and saves and pushes the model to the Hub. If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Once you've made all your changes or you're okay with the default configuration, you're ready to launch the training script! 🚀
<hfoptionsid="training-inference">
<hfoptionid="PyTorch">
Let's train on the [Naruto BLIP captions](https://huggingface.co/datasets/lambdalabs/naruto-blip-captions) dataset to generate your own Naruto characters. Set the environment variables `MODEL_NAME` and `dataset_name` to the model and the dataset (either from the Hub or a local path). If you're training on more than one GPU, add the `--multi_gpu` parameter to the `accelerate launch` command.
<Tip>
To train on a local dataset, set the `TRAIN_DIR` and `OUTPUT_DIR` environment variables to the path of the dataset and where to save the model to.
Training with Flax can be faster on TPUs and GPUs thanks to [@duongna211](https://github.com/duongna21). Flax is more efficient on a TPU, but GPU performance is also great.
Set the environment variables `MODEL_NAME` and `dataset_name` to the model and the dataset (either from the Hub or a local path).
<Tip>
To train on a local dataset, set the `TRAIN_DIR` and `OUTPUT_DIR` environment variables to the path of the dataset and where to save the model to.
Congratulations on training your own text-to-image model! To learn more about how to use your new model, the following guides may be helpful:
- Learn how to [load LoRA weights](../using-diffusers/loading_adapters#LoRA) for inference if you trained your model with LoRA.
- Learn more about how certain parameters like guidance scale or techniques such as prompt weighting can help you control inference in the [Text-to-image](../using-diffusers/conditional_image_generation) task guide.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Textual Inversion
[Textual Inversion](https://hf.co/papers/2208.01618) is a training technique for personalizing image generation models with just a few example images of what you want it to learn. This technique works by learning and updating the text embeddings (the new embeddings are tied to a special word you must use in the prompt) to match the example images you provide.
If you're training on a GPU with limited vRAM, you should try enabling the `gradient_checkpointing` and `mixed_precision` parameters in the training command. You can also reduce your memory footprint by using memory-efficient attention with [xFormers](../optimization/xformers). JAX/Flax training is also supported for efficient training on TPUs and GPUs, but it doesn't support gradient checkpointing or xFormers. With the same configuration and setup as PyTorch, the Flax training script should be at least ~70% faster!
This guide will explore the [textual_inversion.py](https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py) script to help you become more familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Navigate to the example folder with the training script and install the required dependencies for the script you're using:
<hfoptionsid="installation">
<hfoptionid="PyTorch">
```bash
cd examples/textual_inversion
pip install-r requirements.txt
```
</hfoption>
<hfoptionid="Flax">
```bash
cd examples/textual_inversion
pip install-r requirements_flax.txt
```
</hfoption>
</hfoptions>
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell, like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
<Tip>
The following sections highlight parts of the training script that are important for understanding how to modify it, but it doesn't cover every aspect of the script in detail. If you're interested in learning more, feel free to read through the [script](https://github.com/huggingface/diffusers/blob/main/examples/textual_inversion/textual_inversion.py) and let us know if you have any questions or concerns.
</Tip>
## Script parameters
The training script has many parameters to help you tailor the training run to your needs. All of the parameters and their descriptions are listed in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/839c2a5ece0af4e75530cb520d77bc7ed8acf474/examples/textual_inversion/textual_inversion.py#L176) function. Where applicable, Diffusers provides default values for each parameter such as the training batch size and learning rate, but feel free to change these values in the training command if you'd like.
For example, to increase the number of gradient accumulation steps above the default value of 1:
```bash
accelerate launch textual_inversion.py \
--gradient_accumulation_steps=4
```
Some other basic and important parameters to specify include:
-`--pretrained_model_name_or_path`: the name of the model on the Hub or a local path to the pretrained model
-`--train_data_dir`: path to a folder containing the training dataset (example images)
-`--output_dir`: where to save the trained model
-`--push_to_hub`: whether to push the trained model to the Hub
-`--checkpointing_steps`: frequency of saving a checkpoint as the model trains; this is useful if for some reason training is interrupted, you can continue training from that checkpoint by adding `--resume_from_checkpoint` to your training command
-`--num_vectors`: the number of vectors to learn the embeddings with; increasing this parameter helps the model learn better but it comes with increased training costs
-`--placeholder_token`: the special word to tie the learned embeddings to (you must use the word in your prompt for inference)
-`--initializer_token`: a single-word that roughly describes the object or style you're trying to train on
-`--learnable_property`: whether you're training the model to learn a new "style" (for example, Van Gogh's painting style) or "object" (for example, your dog)
## Training script
Unlike some of the other training scripts, textual_inversion.py has a custom dataset class, [`TextualInversionDataset`](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L487) for creating a dataset. You can customize the image size, placeholder token, interpolation method, whether to crop the image, and more. If you need to change how the dataset is created, you can modify `TextualInversionDataset`.
Next, you'll find the dataset preprocessing code and training loop in the [`main()`](https://github.com/huggingface/diffusers/blob/839c2a5ece0af4e75530cb520d77bc7ed8acf474/examples/textual_inversion/textual_inversion.py#L573) function.
The script starts by loading the [tokenizer](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L616), [scheduler and model](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L622):
The special [placeholder token](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L632) is added next to the tokenizer, and the embedding is readjusted to account for the new token.
Then, the script [creates a dataset](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L716) from the `TextualInversionDataset`:
Finally, the [training loop](https://github.com/huggingface/diffusers/blob/b81c69e489aad3a0ba73798c459a33990dc4379c/examples/textual_inversion/textual_inversion.py#L784) handles everything else from predicting the noisy residual to updating the embedding weights of the special placeholder token.
If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Once you've made all your changes or you're okay with the default configuration, you're ready to launch the training script! 🚀
For this guide, you'll download some images of a [cat toy](https://huggingface.co/datasets/diffusers/cat_toy_example) and store them in a directory. But remember, you can create and use your own dataset if you want (see the [Create a dataset for training](create_dataset) guide).
Set the environment variable `MODEL_NAME` to a model id on the Hub or a path to a local model, and `DATA_DIR` to the path where you just downloaded the cat images to. The script creates and saves the following files to your repository:
-`learned_embeds.bin`: the learned embedding vectors corresponding to your example images
-`token_identifier.txt`: the special placeholder token
-`type_of_concept.txt`: the type of concept you're training on (either "object" or "style")
<Tipwarning={true}>
A full training run takes ~1 hour on a single V100 GPU.
</Tip>
One more thing before you launch the script. If you're interested in following along with the training process, you can periodically save generated images as training progresses. Add the following parameters to the training command:
Flax doesn't support the [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] method, but the textual_inversion_flax.py script [saves](https://github.com/huggingface/diffusers/blob/c0f058265161178f2a88849e92b37ffdc81f1dcc/examples/textual_inversion/textual_inversion_flax.py#L636C2-L636C2) the learned embeddings as a part of the model after training. This means you can use the model for inference like any other Flax model:
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Unconditional image generation
Unconditional image generation models are not conditioned on text or images during training. It only generates images that resemble its training data distribution.
This guide will explore the [train_unconditional.py](https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/train_unconditional.py) training script to help you become familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Then navigate to the example folder containing the training script and install the required dependencies:
```bash
cd examples/unconditional_image_generation
pip install-r requirements.txt
```
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
## Script parameters
<Tip>
The following sections highlight parts of the training script that are important for understanding how to modify it, but it doesn't cover every aspect of the script in detail. If you're interested in learning more, feel free to read through the [script](https://github.com/huggingface/diffusers/blob/main/examples/unconditional_image_generation/train_unconditional.py) and let us know if you have any questions or concerns.
</Tip>
The training script provides many parameters to help you customize your training run. All of the parameters and their descriptions are found in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L55) function. It provides default values for each parameter, such as the training batch size and learning rate, but you can also set your own values in the training command if you'd like.
For example, to speedup training with mixed precision using the bf16 format, add the `--mixed_precision` parameter to the training command:
```bash
accelerate launch train_unconditional.py \
--mixed_precision="bf16"
```
Some basic and important parameters to specify include:
-`--dataset_name`: the name of the dataset on the Hub or a local path to the dataset to train on
-`--output_dir`: where to save the trained model
-`--push_to_hub`: whether to push the trained model to the Hub
-`--checkpointing_steps`: frequency of saving a checkpoint as the model trains; this is useful if training is interrupted, you can continue training from that checkpoint by adding `--resume_from_checkpoint` to your training command
Bring your dataset, and let the training script handle everything else!
## Training script
The code for preprocessing the dataset and the training loop is found in the [`main()`](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L275) function. If you need to adapt the training script, this is where you'll need to make your changes.
The `train_unconditional` script [initializes a `UNet2DModel`](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L356) if you don't provide a model configuration. You can configure the UNet here if you'd like:
```py
model=UNet2DModel(
sample_size=args.resolution,
in_channels=3,
out_channels=3,
layers_per_block=2,
block_out_channels=(128,128,256,256,512,512),
down_block_types=(
"DownBlock2D",
"DownBlock2D",
"DownBlock2D",
"DownBlock2D",
"AttnDownBlock2D",
"DownBlock2D",
),
up_block_types=(
"UpBlock2D",
"AttnUpBlock2D",
"UpBlock2D",
"UpBlock2D",
"UpBlock2D",
"UpBlock2D",
),
)
```
Next, the script initializes a [scheduler](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L418) and [optimizer](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L429):
Then it [loads a dataset](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L451) and you can specify how to [preprocess](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L455) it:
Finally, the [training loop](https://github.com/huggingface/diffusers/blob/096f84b05f9514fae9f185cbec0a4d38fbad9919/examples/unconditional_image_generation/train_unconditional.py#L540) handles everything else such as adding noise to the images, predicting the noise residual, calculating the loss, saving checkpoints at specified steps, and saving and pushing the model to the Hub. If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Once you've made all your changes or you're okay with the default configuration, you're ready to launch the training script! 🚀
<Tipwarning={true}>
A full training run takes 2 hours on 4xV100 GPUs.
</Tip>
<hfoptionsid="launchtraining">
<hfoptionid="single GPU">
```bash
accelerate launch train_unconditional.py \
--dataset_name="huggan/flowers-102-categories"\
--output_dir="ddpm-ema-flowers-64"\
--mixed_precision="fp16"\
--push_to_hub
```
</hfoption>
<hfoptionid="multi-GPU">
If you're training with more than one GPU, add the `--multi_gpu` parameter to the training command:
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Wuerstchen
The [Wuerstchen](https://hf.co/papers/2306.00637) model drastically reduces computational costs by compressing the latent space by 42x, without compromising image quality and accelerating inference. During training, Wuerstchen uses two models (VQGAN + autoencoder) to compress the latents, and then a third model (text-conditioned latent diffusion model) is conditioned on this highly compressed space to generate an image.
To fit the prior model into GPU memory and to speedup training, try enabling `gradient_accumulation_steps`, `gradient_checkpointing`, and `mixed_precision` respectively.
This guide explores the [train_text_to_image_prior.py](https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/train_text_to_image_prior.py) script to help you become more familiar with it, and how you can adapt it for your own use-case.
Before running the script, make sure you install the library from source:
Then navigate to the example folder containing the training script and install the required dependencies for the script you're using:
```bash
cd examples/wuerstchen/text_to_image
pip install-r requirements.txt
```
<Tip>
🤗 Accelerate is a library for helping you train on multiple GPUs/TPUs or with mixed-precision. It'll automatically configure your training setup based on your hardware and environment. Take a look at the 🤗 Accelerate [Quick tour](https://huggingface.co/docs/accelerate/quicktour) to learn more.
</Tip>
Initialize an 🤗 Accelerate environment:
```bash
accelerate config
```
To setup a default 🤗 Accelerate environment without choosing any configurations:
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell, like a notebook, you can use:
```py
fromaccelerate.utilsimportwrite_basic_config
write_basic_config()
```
Lastly, if you want to train a model on your own dataset, take a look at the [Create a dataset for training](create_dataset) guide to learn how to create a dataset that works with the training script.
<Tip>
The following sections highlight parts of the training scripts that are important for understanding how to modify it, but it doesn't cover every aspect of the [script](https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/train_text_to_image_prior.py) in detail. If you're interested in learning more, feel free to read through the scripts and let us know if you have any questions or concerns.
</Tip>
## Script parameters
The training scripts provides many parameters to help you customize your training run. All of the parameters and their descriptions are found in the [`parse_args()`](https://github.com/huggingface/diffusers/blob/6e68c71503682c8693cb5b06a4da4911dfd655ee/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L192) function. It provides default values for each parameter, such as the training batch size and learning rate, but you can also set your own values in the training command if you'd like.
For example, to speedup training with mixed precision using the fp16 format, add the `--mixed_precision` parameter to the training command:
```bash
accelerate launch train_text_to_image_prior.py \
--mixed_precision="fp16"
```
Most of the parameters are identical to the parameters in the [Text-to-image](text2image#script-parameters) training guide, so let's dive right into the Wuerstchen training script!
## Training script
The training script is also similar to the [Text-to-image](text2image#training-script) training guide, but it's been modified to support Wuerstchen. This guide focuses on the code that is unique to the Wuerstchen training script.
The [`main()`](https://github.com/huggingface/diffusers/blob/6e68c71503682c8693cb5b06a4da4911dfd655ee/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L441) function starts by initializing the image encoder - an [EfficientNet](https://github.com/huggingface/diffusers/blob/main/examples/wuerstchen/text_to_image/modeling_efficient_net_encoder.py) - in addition to the usual scheduler and tokenizer.
Next, you'll apply some [transforms](https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L656) to the images and [tokenize](https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L637) the captions:
Finally, the [training loop](https://github.com/huggingface/diffusers/blob/65ef7a0c5c594b4f84092e328fbdd73183613b30/examples/wuerstchen/text_to_image/train_text_to_image_prior.py#L656) handles compressing the images to latent space with the `EfficientNetEncoder`, adding noise to the latents, and predicting the noise residual with the [`WuerstchenPrior`] model.
If you want to learn more about how the training loop works, check out the [Understanding pipelines, models and schedulers](../using-diffusers/write_own_pipeline) tutorial which breaks down the basic pattern of the denoising process.
## Launch the script
Once you’ve made all your changes or you’re okay with the default configuration, you’re ready to launch the training script! 🚀
Set the `DATASET_NAME` environment variable to the dataset name from the Hub. This guide uses the [Naruto BLIP captions](https://huggingface.co/datasets/lambdalabs/naruto-blip-captions) dataset, but you can create and train on your own datasets as well (see the [Create a dataset for training](create_dataset) guide).
<Tip>
To monitor training progress with Weights & Biases, add the `--report_to=wandb` parameter to the training command. You’ll also need to add the `--validation_prompt` to the training command to keep track of results. This can be really useful for debugging the model and viewing intermediate results.
Congratulations on training a Wuerstchen model! To learn more about how to use your new model, the following may be helpful:
- Take a look at the [Wuerstchen](../api/pipelines/wuerstchen#text-to-image-generation) API documentation to learn more about how to use the pipeline for text-to-image generation and its limitations.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# AutoPipeline
Diffusers provides many pipelines for basic tasks like generating images, videos, audio, and inpainting. On top of these, there are specialized pipelines for adapters and features like upscaling, super-resolution, and more. Different pipeline classes can even use the same checkpoint because they share the same pretrained model! With so many different pipelines, it can be overwhelming to know which pipeline class to use.
The [AutoPipeline](../api/pipelines/auto_pipeline) class is designed to simplify the variety of pipelines in Diffusers. It is a generic *task-first* pipeline that lets you focus on a task ([`AutoPipelineForText2Image`], [`AutoPipelineForImage2Image`], and [`AutoPipelineForInpainting`]) without needing to know the specific pipeline class. The [AutoPipeline](../api/pipelines/auto_pipeline) automatically detects the correct pipeline class to use.
For example, let's use the [dreamlike-art/dreamlike-photoreal-2.0](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0) checkpoint.
Under the hood, [AutoPipeline](../api/pipelines/auto_pipeline):
1. Detects a `"stable-diffusion"` class from the [model_index.json](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0/blob/main/model_index.json) file.
2. Depending on the task you're interested in, it loads the [`StableDiffusionPipeline`], [`StableDiffusionImg2ImgPipeline`], or [`StableDiffusionInpaintPipeline`]. Any parameter (`strength`, `num_inference_steps`, etc.) you would pass to these specific pipelines can also be passed to the [AutoPipeline](../api/pipelines/auto_pipeline).
Notice how the [dreamlike-art/dreamlike-photoreal-2.0](https://hf.co/dreamlike-art/dreamlike-photoreal-2.0) checkpoint is used for both text-to-image and image-to-image tasks? To save memory and avoid loading the checkpoint twice, use the [`~DiffusionPipeline.from_pipe`] method.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
[[open-in-colab]]
# Train a diffusion model
Unconditional image generation is a popular application of diffusion models that generates images that look like those in the dataset used for training. Typically, the best results are obtained from finetuning a pretrained model on a specific dataset. You can find many of these checkpoints on the [Hub](https://huggingface.co/search/full-text?q=unconditional-image-generation&type=model), but if you can't find one you like, you can always train your own!
This tutorial will teach you how to train a [`UNet2DModel`] from scratch on a subset of the [Smithsonian Butterflies](https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset) dataset to generate your own 🦋 butterflies 🦋.
<Tip>
💡 This training tutorial is based on the [Training with 🧨 Diffusers](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) notebook. For additional details and context about diffusion models like how they work, check out the notebook!
</Tip>
Before you begin, make sure you have 🤗 Datasets installed to load and preprocess image datasets, and 🤗 Accelerate, to simplify training on any number of GPUs. The following command will also install [TensorBoard](https://www.tensorflow.org/tensorboard) to visualize training metrics (you can also use [Weights & Biases](https://docs.wandb.ai/) to track your training).
```py
# uncomment to install the necessary libraries in Colab
#!pip install diffusers[training]
```
We encourage you to share your model with the community, and in order to do that, you'll need to login to your Hugging Face account (create one [here](https://hf.co/join) if you don't already have one!). You can login from a notebook and enter your token when prompted. Make sure your token has the write role.
```py
>>>fromhuggingface_hubimportnotebook_login
>>>notebook_login()
```
Or login in from the terminal:
```bash
huggingface-cli login
```
Since the model checkpoints are quite large, install [Git-LFS](https://git-lfs.com/) to version these large files:
```bash
!sudo apt -qqinstall git-lfs
!git config --global credential.helper store
```
## Training configuration
For convenience, create a `TrainingConfig` class containing the training hyperparameters (feel free to adjust them):
```py
>>>fromdataclassesimportdataclass
>>>@dataclass
...classTrainingConfig:
...image_size=128# the generated image resolution
...train_batch_size=16
...eval_batch_size=16# how many images to sample during evaluation
...num_epochs=50
...gradient_accumulation_steps=1
...learning_rate=1e-4
...lr_warmup_steps=500
...save_image_epochs=10
...save_model_epochs=30
...mixed_precision="fp16"# `no` for float32, `fp16` for automatic mixed precision
...output_dir="ddpm-butterflies-128"# the model name locally and on the HF Hub
...push_to_hub=True# whether to upload the saved model to the HF Hub
...hub_model_id="<your-username>/<my-awesome-model>"# the name of the repository to create on the HF Hub
...hub_private_repo=None
...overwrite_output_dir=True# overwrite the old model when re-running the notebook
...seed=0
>>>config=TrainingConfig()
```
## Load the dataset
You can easily load the [Smithsonian Butterflies](https://huggingface.co/datasets/huggan/smithsonian_butterflies_subset) dataset with the 🤗 Datasets library:
💡 You can find additional datasets from the [HugGan Community Event](https://huggingface.co/huggan) or you can use your own dataset by creating a local [`ImageFolder`](https://huggingface.co/docs/datasets/image_dataset#imagefolder). Set `config.dataset_name` to the repository id of the dataset if it is from the HugGan Community Event, or `imagefolder` if you're using your own images.
</Tip>
🤗 Datasets uses the [`~datasets.Image`] feature to automatically decode the image data and load it as a [`PIL.Image`](https://pillow.readthedocs.io/en/stable/reference/Image.html) which we can visualize:
Feel free to visualize the images again to confirm that they've been resized. Now you're ready to wrap the dataset in a [DataLoader](https://pytorch.org/docs/stable/data#torch.utils.data.DataLoader) for training!
Great! Next, you'll need a scheduler to add some noise to the image.
## Create a scheduler
The scheduler behaves differently depending on whether you're using the model for training or inference. During inference, the scheduler generates image from the noise. During training, the scheduler takes a model output - or a sample - from a specific point in the diffusion process and applies noise to the image according to a *noise schedule* and an *update rule*.
Let's take a look at the [`DDPMScheduler`] and use the `add_noise` method to add some random noise to the `sample_image` from before:
Then, you'll need a way to evaluate the model. For evaluation, you can use the [`DDPMPipeline`] to generate a batch of sample images and save it as a grid:
```py
>>>fromdiffusersimportDDPMPipeline
>>>fromdiffusers.utilsimportmake_image_grid
>>>importos
>>>defevaluate(config,epoch,pipeline):
...# Sample some images from random noise (this is the backward diffusion process).
...# The default pipeline output type is `List[PIL.Image]`
...images=pipeline(
...batch_size=config.eval_batch_size,
...generator=torch.Generator(device='cpu').manual_seed(config.seed),# Use a separate torch generator to avoid rewinding the random state of the main training loop
Now you can wrap all these components together in a training loop with 🤗 Accelerate for easy TensorBoard logging, gradient accumulation, and mixed precision training. To upload the model to the Hub, write a function to get your repository name and information and then push it to the Hub.
<Tip>
💡 The training loop below may look intimidating and long, but it'll be worth it later when you launch your training in just one line of code! If you can't wait and want to start generating images, feel free to copy and run the code below. You can always come back and examine the training loop more closely later, like when you're waiting for your model to finish training. 🤗
Phew, that was quite a bit of code! But you're finally ready to launch the training with 🤗 Accelerate's [`~accelerate.notebook_launcher`] function. Pass the function the training loop, all the training arguments, and the number of processes (you can change this value to the number of GPUs available to you) to use for training:
Unconditional image generation is one example of a task that can be trained. You can explore other tasks and training techniques by visiting the [🧨 Diffusers Training Examples](../training/overview) page. Here are some examples of what you can learn:
*[Textual Inversion](../training/text_inversion), an algorithm that teaches a model a specific visual concept and integrates it into the generated image.
*[DreamBooth](../training/dreambooth), a technique for generating personalized images of a subject given several input images of the subject.
*[Guide](../training/text2image) to finetuning a Stable Diffusion model on your own dataset.
*[Guide](../training/lora) to using LoRA, a memory-efficient technique for finetuning really large models faster.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Accelerate inference of text-to-image diffusion models
Diffusion models are slower than their GAN counterparts because of the iterative and sequential reverse diffusion process. There are several techniques that can address this limitation such as progressive timestep distillation ([LCM LoRA](../using-diffusers/inference_with_lcm_lora)), model compression ([SSD-1B](https://huggingface.co/segmind/SSD-1B)), and reusing adjacent features of the denoiser ([DeepCache](../optimization/deepcache)).
However, you don't necessarily need to use these techniques to speed up inference. With PyTorch 2 alone, you can accelerate the inference latency of text-to-image diffusion pipelines by up to 3x. This tutorial will show you how to progressively apply the optimizations found in PyTorch 2 to reduce inference latency. You'll use the [Stable Diffusion XL (SDXL)](../using-diffusers/sdxl) pipeline in this tutorial, but these techniques are applicable to other text-to-image diffusion pipelines too.
Make sure you're using the latest version of Diffusers:
```bash
pip install-U diffusers
```
Then upgrade the other required libraries too:
```bash
pip install-U transformers accelerate peft
```
Install [PyTorch nightly](https://pytorch.org/) to benefit from the latest and fastest kernels:
> The results reported below are from a 80GB 400W A100 with its clock rate set to the maximum.
> If you're interested in the full benchmarking code, take a look at [huggingface/diffusion-fast](https://github.com/huggingface/diffusion-fast).
## Baseline
Let's start with a baseline. Disable reduced precision and the [`scaled_dot_product_attention` (SDPA)](../optimization/torch2.0#scaled-dot-product-attention) function which is automatically used by Diffusers:
```python
fromdiffusersimportStableDiffusionXLPipeline
# Load the pipeline in full-precision and place its model components on CUDA.
pipe=StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0"
).to("cuda")
# Run the attention ops without SDPA.
pipe.unet.set_default_attn_processor()
pipe.vae.set_default_attn_processor()
prompt="Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
Enable the first optimization, reduced precision or more specifically bfloat16. There are several benefits of using reduced precision:
* Using a reduced numerical precision (such as float16 or bfloat16) for inference doesn’t affect the generation quality but significantly improves latency.
* The benefits of using bfloat16 compared to float16 are hardware dependent, but modern GPUs tend to favor bfloat16.
* bfloat16 is much more resilient when used with quantization compared to float16, but more recent versions of the quantization library ([torchao](https://github.com/pytorch-labs/ao)) we used don't have numerical issues with float16.
In our later experiments with float16, recent versions of torchao do not incur numerical problems from float16.
</Tip>
Take a look at the [Speed up inference](../optimization/fp16) guide to learn more about running inference with reduced precision.
## SDPA
Attention blocks are intensive to run. But with PyTorch's [`scaled_dot_product_attention`](../optimization/torch2.0#scaled-dot-product-attention) function, it is a lot more efficient. This function is used by default in Diffusers so you don't need to make any changes to the code.
PyTorch 2 includes `torch.compile` which uses fast and optimized kernels. In Diffusers, the UNet and VAE are usually compiled because these are the most compute-intensive modules. First, configure a few compiler flags (refer to the [full list](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/config.py) for more options):
`torch.compile` offers different backends and modes. For maximum inference speed, use "max-autotune" for the inductor backend. “max-autotune” uses CUDA graphs and optimizes the compilation graph specifically for latency. CUDA graphs greatly reduces the overhead of launching GPU operations by using a mechanism to launch multiple GPU operations through a single CPU operation.
Using SDPA attention and compiling both the UNet and VAE cuts the latency from 3.31 seconds to 2.54 seconds.
> From PyTorch 2.3.1, you can control the caching behavior of `torch.compile()`. This is particularly beneficial for compilation modes like `"max-autotune"` which performs a grid-search over several compilation flags to find the optimal configuration. Learn more in the [Compile Time Caching in torch.compile](https://pytorch.org/tutorials/recipes/torch_compile_caching_tutorial.html) tutorial.
### Prevent graph breaks
Specifying `fullgraph=True` ensures there are no graph breaks in the underlying model to take full advantage of `torch.compile` without any performance degradation. For the UNet and VAE, this means changing how you access the return variables.
During the iterative reverse diffusion process, the `step()` function is [called](https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py#L1228) on the scheduler each time after the denoiser predicts the less noisy latent embeddings. Inside `step()`, the `sigmas` variable is [indexed](https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/schedulers/scheduling_euler_discrete.py#L476) which when placed on the GPU, causes a communication sync between the CPU and GPU. This introduces latency and it becomes more evident when the denoiser has already been compiled.
But if the `sigmas` array always [stays on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240), the CPU and GPU sync doesn’t occur and you don't get any latency. In general, any CPU and GPU communication sync should be none or be kept to a bare minimum because it can impact inference latency.
## Combine the attention block's projection matrices
The UNet and VAE in SDXL use Transformer-like blocks which consists of attention blocks and feed-forward blocks.
In an attention block, the input is projected into three sub-spaces using three different projection matrices – Q, K, and V. These projections are performed separately on the input. But we can horizontally combine the projection matrices into a single matrix and perform the projection in one step. This increases the size of the matrix multiplications of the input projections and improves the impact of quantization.
You can combine the projection matrices with just a single line of code:
```python
pipe.fuse_qkv_projections()
```
This provides a minor improvement from 2.54 seconds to 2.52 seconds.
Support for [`~StableDiffusionXLPipeline.fuse_qkv_projections`] is limited and experimental. It's not available for many non-Stable Diffusion pipelines such as [Kandinsky](../using-diffusers/kandinsky). You can refer to this [PR](https://github.com/huggingface/diffusers/pull/6179) to get an idea about how to enable this for the other pipelines.
</Tip>
## Dynamic quantization
You can also use the ultra-lightweight PyTorch quantization library, [torchao](https://github.com/pytorch-labs/ao)(commit SHA `54bcd5a10d0abbe7b0c045052029257099f83fd9`), to apply [dynamic int8 quantization](https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html) to the UNet and VAE. Quantization adds additional conversion overhead to the model that is hopefully made up for by faster matmuls (dynamic quantization). If the matmuls are too small, these techniques may degrade performance.
Certain linear layers in the UNet and VAE don’t benefit from dynamic int8 quantization. You can filter out those layers with the [`dynamic_quant_filter_fn`](https://github.com/huggingface/diffusion-fast/blob/0f169640b1db106fe6a479f78c1ed3bfaeba3386/utils/pipeline_utils.py#L16) shown below.
Since dynamic quantization is only limited to the linear layers, convert the appropriate pointwise convolution layers into linear layers to maximize its benefit.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Working with big models
A modern diffusion model, like [Stable Diffusion XL (SDXL)](../using-diffusers/sdxl), is not just a single model, but a collection of multiple models. SDXL has four different model-level components:
* A variational autoencoder (VAE)
* Two text encoders
* A UNet for denoising
Usually, the text encoders and the denoiser are much larger compared to the VAE.
As models get bigger and better, it’s possible your model is so big that even a single copy won’t fit in memory. But that doesn’t mean it can’t be loaded. If you have more than one GPU, there is more memory available to store your model. In this case, it’s better to split your model checkpoint into several smaller *checkpoint shards*.
When a text encoder checkpoint has multiple shards, like [T5-xxl for SD3](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers/tree/main/text_encoder_3), it is automatically handled by the [Transformers](https://huggingface.co/docs/transformers/index) library as it is a required dependency of Diffusers when using the [`StableDiffusion3Pipeline`]. More specifically, Transformers will automatically handle the loading of multiple shards within the requested model class and get it ready so that inference can be performed.
The denoiser checkpoint can also have multiple shards and supports inference thanks to the [Accelerate](https://huggingface.co/docs/accelerate/index) library.
> [!TIP]
> Refer to the [Handling big models for inference](https://huggingface.co/docs/accelerate/main/en/concept_guides/big_model_inference) guide for general guidance when working with big models that are hard to fit into memory.
For example, let's save a sharded checkpoint for the [SDXL UNet](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/unet):
The size of the fp32 variant of the SDXL UNet checkpoint is ~10.4GB. Set the `max_shard_size` parameter to 5GB to create 3 shards. After saving, you can load them in [`StableDiffusionXLPipeline`]:
image=pipeline("a cute dog running on the grass",num_inference_steps=30).images[0]
image.save("dog.png")
```
If placing all the model-level components on the GPU at once is not feasible, use [`~DiffusionPipeline.enable_model_cpu_offload`] to help you:
```diff
- pipeline.to("cuda")
+ pipeline.enable_model_cpu_offload()
```
In general, we recommend sharding when a checkpoint is more than 5GB (in fp32).
## Device placement
On distributed setups, you can run inference across multiple GPUs with Accelerate.
> [!WARNING]
> This feature is experimental and its APIs might change in the future.
With Accelerate, you can use the `device_map` to determine how to distribute the models of a pipeline across multiple devices. This is useful in situations where you have more than one GPU.
For example, if you have two 8GB GPUs, then using [`~DiffusionPipeline.enable_model_cpu_offload`] may not work so well because:
* it only works on a single GPU
* a single model might not fit on a single GPU ([`~DiffusionPipeline.enable_sequential_cpu_offload`] might work but it will be extremely slow and it is also limited to a single GPU)
To make use of both GPUs, you can use the "balanced" device placement strategy which splits the models across all available GPUs.
> [!WARNING]
> Only the "balanced" strategy is supported at the moment, and we plan to support additional mapping strategies in the future.
You can also pass a dictionary to enforce the maximum GPU memory that can be used on each device:
```diff
from diffusers import DiffusionPipeline
import torch
max_memory = {0:"1GB", 1:"1GB"}
pipeline = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16,
use_safetensors=True,
device_map="balanced",
+ max_memory=max_memory
)
image = pipeline("a dog").images[0]
image
```
If a device is not present in `max_memory`, then it will be completely ignored and will not participate in the device placement.
By default, Diffusers uses the maximum memory of all devices. If the models don't fit on the GPUs, they are offloaded to the CPU. If the CPU doesn't have enough memory, then you might see an error. In that case, you could defer to using [`~DiffusionPipeline.enable_sequential_cpu_offload`] and [`~DiffusionPipeline.enable_model_cpu_offload`].
Call [`~DiffusionPipeline.reset_device_map`] to reset the `device_map` of a pipeline. This is also necessary if you want to use methods like `to()`, [`~DiffusionPipeline.enable_sequential_cpu_offload`], and [`~DiffusionPipeline.enable_model_cpu_offload`] on a pipeline that was device-mapped.
```py
pipeline.reset_device_map()
```
Once a pipeline has been device-mapped, you can also access its device map via `hf_device_map`:
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Overview
Welcome to 🧨 Diffusers! If you're new to diffusion models and generative AI, and want to learn more, then you've come to the right place. These beginner-friendly tutorials are designed to provide a gentle introduction to diffusion models and help you understand the library fundamentals - the core components and how 🧨 Diffusers is meant to be used.
You'll learn how to use a pipeline for inference to rapidly generate things, and then deconstruct that pipeline to really understand how to use the library as a modular toolbox for building your own diffusion systems. In the next lesson, you'll learn how to train your own diffusion model to generate what you want.
After completing the tutorials, you'll have gained the necessary skills to start exploring the library on your own and see how to use it for your own projects and applications.
Feel free to join our community on [Discord](https://discord.com/invite/JfAtkvEtRb) or the [forums](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) to connect and collaborate with other users and developers!
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
[[open-in-colab]]
# Load LoRAs for inference
There are many adapter types (with [LoRAs](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) being the most popular) trained in different styles to achieve different effects. You can even combine multiple adapters to create new and unique images.
In this tutorial, you'll learn how to easily load and manage adapters for inference with the 🤗 [PEFT](https://huggingface.co/docs/peft/index) integration in 🤗 Diffusers. You'll use LoRA as the main adapter technique, so you'll see the terms LoRA and adapter used interchangeably.
Next, load a [CiroN2022/toy-face](https://huggingface.co/CiroN2022/toy-face) adapter with the [`~diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] method. With the 🤗 PEFT integration, you can assign a specific `adapter_name` to the checkpoint, which lets you easily switch between different LoRA checkpoints. Let's call this adapter `"toy"`.
With the `adapter_name` parameter, it is really easy to use another adapter for inference! Load the [nerijs/pixel-art-xl](https://huggingface.co/nerijs/pixel-art-xl) adapter that has been fine-tuned to generate pixel art images and call it `"pixel"`.
The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~PeftAdapterMixin.set_adapters`] method:
By default, if the most up-to-date versions of PEFT and Transformers are detected, `low_cpu_mem_usage` is set to `True` to speed up the loading time of LoRA checkpoints.
</Tip>
## Merge adapters
You can also merge different adapter checkpoints for inference to blend their styles together.
Once again, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
LoRA checkpoints in the diffusion community are almost always obtained with [DreamBooth](https://huggingface.co/docs/diffusers/main/en/training/dreambooth). DreamBooth training often relies on "trigger" words in the input text prompts in order for the generation results to look as expected. When you combine multiple LoRA checkpoints, it's important to ensure the trigger words for the corresponding LoRA checkpoints are present in the input text prompts.
</Tip>
Remember to use the trigger words for [CiroN2022/toy-face](https://hf.co/CiroN2022/toy-face) and [nerijs/pixel-art-xl](https://hf.co/nerijs/pixel-art-xl)(these are found in their repositories) in the prompt to generate an image.
```python
prompt="toy_face of a hacker with a hoodie, pixel art"
Impressive! As you can see, the model generated an image that mixed the characteristics of both adapters.
> [!TIP]
> Through its PEFT integration, Diffusers also offers more efficient merging methods which you can learn about in the [Merge LoRAs](../using-diffusers/merge_loras) guide!
To return to only using one adapter, use the [`~PeftAdapterMixin.set_adapters`] method to activate the `"toy"` adapter:
For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~PeftAdapterMixin.set_adapters`].
For example, here's how you can turn on the adapter for the `down` parts, but turn it off for the `mid` and `up` parts:
```python
pipe.enable_lora()# enable lora again, after we disabled it above
prompt="toy_face of a hacker with a hoodie, pixel art"
This is a really powerful feature. You can use it to control the adapter strengths down to per-transformer level. And you can even use it for multiple adapters.
```python
adapter_weight_scales_toy=0.5
adapter_weight_scales_pixel={
"unet":{
"down":0.9,# all transformers in the down-part will use scale 0.9
# "mid" # because, in this example, "mid" is not given, all transformers in the mid part will use the default scale 1.0
"up":{
"block_0":0.6,# all 3 transformers in the 0th block in the up-part will use scale 0.6
"block_1":[0.4,0.8,1.0],# the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.StableDiffusionLoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:
```py
active_adapters=pipe.get_active_adapters()
active_adapters
["toy","pixel"]
```
You can also get the active adapters of each pipeline component with [`~diffusers.loaders.StableDiffusionLoraLoaderMixin.get_list_adapters`]:
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Pipeline callbacks
The denoising loop of a pipeline can be modified with custom defined functions using the `callback_on_step_end` parameter. The callback function is executed at the end of each step, and modifies the pipeline attributes and variables for the next step. This is really useful for *dynamically* adjusting certain pipeline attributes or modifying tensor variables. This versatility allows for interesting use cases such as changing the prompt embeddings at each timestep, assigning different weights to the prompt embeddings, and editing the guidance scale. With callbacks, you can implement new features without modifying the underlying code!
> [!TIP]
> 🤗 Diffusers currently only supports `callback_on_step_end`, but feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) if you have a cool use-case and require a callback function with a different execution point!
This guide will demonstrate how callbacks work by a few features you can implement with them.
## Official callbacks
We provide a list of callbacks you can plug into an existing pipeline and modify the denoising loop. This is the current list of official callbacks:
-`SDCFGCutoffCallback`: Disables the CFG after a certain number of steps for all SD 1.5 pipelines, including text-to-image, image-to-image, inpaint, and controlnet.
-`SDXLCFGCutoffCallback`: Disables the CFG after a certain number of steps for all SDXL pipelines, including text-to-image, image-to-image, inpaint, and controlnet.
-`IPAdapterScaleCutoffCallback`: Disables the IP Adapter after a certain number of steps for all pipelines supporting IP-Adapter.
> [!TIP]
> If you want to add a new official callback, feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) or [submit a PR](https://huggingface.co/docs/diffusers/main/en/conceptual/contribution#how-to-open-a-pr).
To set up a callback, you need to specify the number of denoising steps after which the callback comes into effect. You can do so by using either one of these two arguments
-`cutoff_step_ratio`: Float number with the ratio of the steps.
-`cutoff_step_index`: Integer number with the exact number of the step.
<imgclass="rounded-xl"src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/without_cfg_callback.png"alt="generated image of a sports car at the road"/>
<imgclass="rounded-xl"src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/with_cfg_callback.png"alt="generated image of a sports car at the road with cfg callback"/>
Dynamic classifier-free guidance (CFG) is a feature that allows you to disable CFG after a certain number of inference steps which can help you save compute with minimal cost to performance. The callback function for this should have the following arguments:
-`pipeline` (or the pipeline instance) provides access to important properties such as `num_timesteps` and `guidance_scale`. You can modify these properties by updating the underlying attributes. For this example, you'll disable CFG by setting `pipeline._guidance_scale=0.0`.
-`step_index` and `timestep` tell you where you are in the denoising loop. Use `step_index` to turn off CFG after reaching 40% of `num_timesteps`.
-`callback_kwargs` is a dict that contains tensor variables you can modify during the denoising loop. It only includes variables specified in the `callback_on_step_end_tensor_inputs` argument, which is passed to the pipeline's `__call__` method. Different pipelines may use different sets of variables, so please check a pipeline's `_callback_tensor_inputs` attribute for the list of variables you can modify. Some common variables include `latents` and `prompt_embeds`. For this function, change the batch size of `prompt_embeds` after setting `guidance_scale=0.0` in order for it to work properly.
Your callback function should look something like this:
> The interruption callback is supported for text-to-image, image-to-image, and inpainting for the [StableDiffusionPipeline](../api/pipelines/stable_diffusion/overview) and [StableDiffusionXLPipeline](../api/pipelines/stable_diffusion/stable_diffusion_xl).
Stopping the diffusion process early is useful when building UIs that work with Diffusers because it allows users to stop the generation process if they're unhappy with the intermediate results. You can incorporate this into your pipeline with a callback.
This callback function should take the following arguments: `pipeline`, `i`, `t`, and `callback_kwargs` (this must be returned). Set the pipeline's `_interrupt` attribute to `True` to stop the diffusion process after a certain number of steps. You are also free to implement your own custom stopping logic inside the callback.
In this example, the diffusion process is stopped after 10 steps even though `num_inference_steps` is set to 50.
> This tip was contributed by [asomoza](https://github.com/asomoza).
Display an image after each generation step by accessing and converting the latents after each step into an image. The latent space is compressed to 128x128, so the images are also 128x128 which is useful for a quick preview.
1. Use the function below to convert the SDXL latents (4 channels) to RGB tensors (3 channels) as explained in the [Explaining the SDXL latent space](https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space) blog post.
3. Pass the `decode_tensors` function to the `callback_on_step_end` parameter to decode the tensors after each step. You also need to specify what you want to modify in the `callback_on_step_end_tensor_inputs` parameter, which in this case are the latents.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# CogVideoX
CogVideoX is a text-to-video generation model focused on creating more coherent videos aligned with a prompt. It achieves this using several methods.
- a 3D variational autoencoder that compresses videos spatially and temporally, improving compression rate and video accuracy.
- an expert transformer block to help align text and video, and a 3D full attention module for capturing and creating spatially and temporally accurate videos.
## Load model checkpoints
Model weights may be stored in separate subfolders on the Hub or locally, in which case, you should use the [`~DiffusionPipeline.from_pretrained`] method.
For text-to-video, pass a text prompt. By default, CogVideoX generates a 720x480 video for the best results.
```py
importtorch
fromdiffusersimportCogVideoXPipeline
fromdiffusers.utilsimportexport_to_video
prompt="An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea."
<imgsrc="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_out.gif"alt="generated image of an astronaut in a jungle"/>
</div>
## Image-to-Video
You'll use the [THUDM/CogVideoX-5b-I2V](https://huggingface.co/THUDM/CogVideoX-5b-I2V) checkpoint for this guide.
prompt="A vast, shimmering ocean flows gracefully under a twilight sky, its waves undulating in a mesmerizing dance of blues and greens. The surface glints with the last rays of the setting sun, casting golden highlights that ripple across the water. Seagulls soar above, their cries blending with the gentle roar of the waves. The horizon stretches infinitely, where the ocean meets the sky in a seamless blend of hues. Close-ups reveal the intricate patterns of the waves, capturing the fluidity and dynamic beauty of the sea in motion."
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Text-to-image
[[open-in-colab]]
When you think of diffusion models, text-to-image is usually one of the first things that come to mind. Text-to-image generates an image from a text description (for example, "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k") which is also known as a *prompt*.
From a very high level, a diffusion model takes a prompt and some random initial noise, and iteratively removes the noise to construct an image. The *denoising* process is guided by the prompt, and once the denoising process ends after a predetermined number of time steps, the image representation is decoded into an image.
<Tip>
Read the [How does Stable Diffusion work?](https://huggingface.co/blog/stable_diffusion#how-does-stable-diffusion-work) blog post to learn more about how a latent diffusion model works.
</Tip>
You can generate images from a prompt in 🤗 Diffusers in two steps:
1. Load a checkpoint into the [`AutoPipelineForText2Image`] class, which automatically detects the appropriate pipeline class to use based on the checkpoint:
The most common text-to-image models are [Stable Diffusion v1.5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5), [Stable Diffusion XL (SDXL)](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0), and [Kandinsky 2.2](https://huggingface.co/kandinsky-community/kandinsky-2-2-decoder). There are also ControlNet models or adapters that can be used with text-to-image models for more direct control in generating images. The results from each model are slightly different because of their architecture and training process, but no matter which model you choose, their usage is more or less the same. Let's use the same prompt for each model and compare their results.
### Stable Diffusion v1.5
[Stable Diffusion v1.5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) is a latent diffusion model initialized from [Stable Diffusion v1-4](https://huggingface.co/CompVis/stable-diffusion-v1-4), and finetuned for 595K steps on 512x512 images from the LAION-Aesthetics V2 dataset. You can use this model like:
image=pipeline("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",generator=generator).images[0]
image
```
### Stable Diffusion XL
SDXL is a much larger version of the previous Stable Diffusion models, and involves a two-stage model process that adds even more details to an image. It also includes some additional *micro-conditionings* to generate high-quality images centered subjects. Take a look at the more comprehensive [SDXL](sdxl) guide to learn more about how to use it. In general, you can use SDXL like:
image=pipeline("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",generator=generator).images[0]
image
```
### Kandinsky 2.2
The Kandinsky model is a bit different from the Stable Diffusion models because it also uses an image prior model to create embeddings that are used to better align text and images in the diffusion model.
image=pipeline("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",generator=generator).images[0]
image
```
### ControlNet
ControlNet models are auxiliary models or adapters that are finetuned on top of text-to-image models, such as [Stable Diffusion v1.5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5). Using ControlNet models in combination with text-to-image models offers diverse options for more explicit control over how to generate an image. With ControlNet, you add an additional conditioning input image to the model. For example, if you provide an image of a human pose (usually represented as multiple keypoints that are connected into a skeleton) as a conditioning input, the model generates an image that follows the pose of the image. Check out the more in-depth [ControlNet](controlnet) guide to learn more about other conditioning inputs and how to use them.
In this example, let's condition the ControlNet with a human pose estimation image. Load the ControlNet model pretrained on human pose estimations:
There are a number of parameters that can be configured in the pipeline that affect how an image is generated. You can change the image's output size, specify a negative prompt to improve image quality, and more. This section dives deeper into how to use these parameters.
### Height and width
The `height` and `width` parameters control the height and width (in pixels) of the generated image. By default, the Stable Diffusion v1.5 model outputs 512x512 images, but you can change this to any size that is a multiple of 8. For example, to create a rectangular image:
Other models may have different default image sizes depending on the image sizes in the training dataset. For example, SDXL's default image size is 1024x1024 and using lower `height` and `width` values may result in lower quality images. Make sure you check the model's API reference first!
</Tip>
### Guidance scale
The `guidance_scale` parameter affects how much the prompt influences image generation. A lower value gives the model "creativity" to generate images that are more loosely related to the prompt. Higher `guidance_scale` values push the model to follow the prompt more closely, and if this value is too high, you may observe some artifacts in the generated image.
Just like how a prompt guides generation, a *negative prompt* steers the model away from things you don't want the model to generate. This is commonly used to improve overall image quality by removing poor or bad image features such as "low resolution" or "bad details". You can also use a negative prompt to remove or modify the content and style of an image.
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html#generator) object enables reproducibility in a pipeline by setting a manual seed. You can use a `Generator` to generate batches of images and iteratively improve on an image generated from a seed as detailed in the [Improve image quality with deterministic generation](reusing_seeds) guide.
You can set a seed and `Generator` as shown below. Creating an image with a `Generator` should return the same result each time instead of randomly generating a new image.
"Astronaut in a jungle, cold color palette, muted colors, detailed, 8k",
generator=generator,
).images[0]
image
```
## Control image generation
There are several ways to exert more control over how an image is generated outside of configuring a pipeline's parameters, such as prompt weighting and ControlNet models.
### Prompt weighting
Prompt weighting is a technique for increasing or decreasing the importance of concepts in a prompt to emphasize or minimize certain features in an image. We recommend using the [Compel](https://github.com/damian0815/compel) library to help you generate the weighted prompt embeddings.
<Tip>
Learn how to create the prompt embeddings in the [Prompt weighting](weighted_prompts) guide. This example focuses on how to use the prompt embeddings in the pipeline.
</Tip>
Once you've created the embeddings, you can pass them to the `prompt_embeds` (and `negative_prompt_embeds` if you're using a negative prompt) parameter in the pipeline.
prompt_embeds=prompt_embeds,# generated from Compel
negative_prompt_embeds=negative_prompt_embeds,# generated from Compel
).images[0]
```
### ControlNet
As you saw in the [ControlNet](#controlnet) section, these models offer a more flexible and accurate way to generate images by incorporating an additional conditioning image input. Each ControlNet model is pretrained on a particular type of conditioning image to generate new images that resemble it. For example, if you take a ControlNet model pretrained on depth maps, you can give the model a depth map as a conditioning input and it'll generate an image that preserves the spatial information in it. This is quicker and easier than specifying the depth information in a prompt. You can even combine multiple conditioning inputs with a [MultiControlNet](controlnet#multicontrolnet)!
There are many types of conditioning inputs you can use, and 🤗 Diffusers supports ControlNet for Stable Diffusion and SDXL models. Take a look at the more comprehensive [ControlNet](controlnet) guide to learn how you can use these models.
## Optimize
Diffusion models are large, and the iterative nature of denoising an image is computationally expensive and intensive. But this doesn't mean you need access to powerful - or even many - GPUs to use them. There are many optimization techniques for running diffusion models on consumer and free-tier resources. For example, you can load model weights in half-precision to save GPU memory and increase speed or offload the entire model to the GPU to save even more memory.
PyTorch 2.0 also supports a more memory-efficient attention mechanism called [*scaled dot product attention*](../optimization/torch2.0#scaled-dot-product-attention) that is automatically enabled if you're using PyTorch 2.0. You can combine this with [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) to speed your code up even more:
For more tips on how to optimize your code to save memory and speed up inference, read the [Memory and speed](../optimization/fp16) and [Torch 2.0](../optimization/torch2.0) guides.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Controlled generation
Controlling outputs generated by diffusion models has been long pursued by the community and is now an active research topic. In many popular diffusion models, subtle changes in inputs, both images and text prompts, can drastically change outputs. In an ideal world we want to be able to control how semantics are preserved and changed.
Most examples of preserving semantics reduce to being able to accurately map a change in input to a change in output. I.e. adding an adjective to a subject in a prompt preserves the entire image, only modifying the changed subject. Or, image variation of a particular subject preserves the subject's pose.
Additionally, there are qualities of generated images that we would like to influence beyond semantic preservation. I.e. in general, we would like our outputs to be of good quality, adhere to a particular style, or be realistic.
We will document some of the techniques `diffusers` supports to control generation of diffusion models. Much is cutting edge research and can be quite nuanced. If something needs clarifying or you have a suggestion, don't hesitate to open a discussion on the [forum](https://discuss.huggingface.co/c/discussion-related-to-httpsgithubcomhuggingfacediffusers/63) or a [GitHub issue](https://github.com/huggingface/diffusers/issues).
We provide a high level explanation of how the generation can be controlled as well as a snippet of the technicals. For more in depth explanations on the technicals, the original papers which are linked from the pipelines are always the best resources.
Depending on the use case, one should choose a technique accordingly. In many cases, these techniques can be combined. For example, one can combine Textual Inversion with SEGA to provide more semantic guidance to the outputs generated using Textual Inversion.
Unless otherwise mentioned, these are techniques that work with existing models and don't require their own weights.
[InstructPix2Pix](../api/pipelines/pix2pix) is fine-tuned from Stable Diffusion to support editing input images. It takes as inputs an image and a prompt describing an edit, and it outputs the edited image.
InstructPix2Pix has been explicitly trained to work well with [InstructGPT](https://openai.com/blog/instruction-following/)-like prompts.
## Pix2Pix Zero
[Paper](https://arxiv.org/abs/2302.03027)
[Pix2Pix Zero](../api/pipelines/pix2pix_zero) allows modifying an image so that one concept or subject is translated to another one while preserving general image semantics.
The denoising process is guided from one conceptual embedding towards another conceptual embedding. The intermediate latents are optimized during the denoising process to push the attention maps towards reference attention maps. The reference attention maps are from the denoising process of the input image and are used to encourage semantic preservation.
Pix2Pix Zero can be used both to edit synthetic images as well as real images.
- To edit synthetic images, one first generates an image given a caption.
Next, we generate image captions for the concept that shall be edited and for the new target concept. We can use a model like [Flan-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5) for this purpose. Then, "mean" prompt embeddings for both the source and target concepts are created via the text encoder. Finally, the pix2pix-zero algorithm is used to edit the synthetic image.
- To edit a real image, one first generates an image caption using a model like [BLIP](https://huggingface.co/docs/transformers/model_doc/blip). Then one applies DDIM inversion on the prompt and image to generate "inverse" latents. Similar to before, "mean" prompt embeddings for both source and target concepts are created and finally the pix2pix-zero algorithm in combination with the "inverse" latents is used to edit the image.
<Tip>
Pix2Pix Zero is the first model that allows "zero-shot" image editing. This means that the model
can edit an image in less than a minute on a consumer GPU as shown [here](../api/pipelines/pix2pix_zero#usage-example).
</Tip>
As mentioned above, Pix2Pix Zero includes optimizing the latents (and not any of the UNet, VAE, or the text encoder) to steer the generation toward a specific concept. This means that the overall
pipeline might require more memory than a standard [StableDiffusionPipeline](../api/pipelines/stable_diffusion/text2img).
<Tip>
An important distinction between methods like InstructPix2Pix and Pix2Pix Zero is that the former
involves fine-tuning the pre-trained weights while the latter does not. This means that you can
apply Pix2Pix Zero to any of the available Stable Diffusion models.
</Tip>
## Attend and Excite
[Paper](https://arxiv.org/abs/2301.13826)
[Attend and Excite](../api/pipelines/attend_and_excite) allows subjects in the prompt to be faithfully represented in the final image.
A set of token indices are given as input, corresponding to the subjects in the prompt that need to be present in the image. During denoising, each token index is guaranteed to have a minimum attention threshold for at least one patch of the image. The intermediate latents are iteratively optimized during the denoising process to strengthen the attention of the most neglected subject token until the attention threshold is passed for all subject tokens.
Like Pix2Pix Zero, Attend and Excite also involves a mini optimization loop (leaving the pre-trained weights untouched) in its pipeline and can require more memory than the usual [StableDiffusionPipeline](../api/pipelines/stable_diffusion/text2img).
## Semantic Guidance (SEGA)
[Paper](https://arxiv.org/abs/2301.12247)
[SEGA](../api/pipelines/semantic_stable_diffusion) allows applying or removing one or more concepts from an image. The strength of the concept can also be controlled. I.e. the smile concept can be used to incrementally increase or decrease the smile of a portrait.
Similar to how classifier free guidance provides guidance via empty prompt inputs, SEGA provides guidance on conceptual prompts. Multiple of these conceptual prompts can be applied simultaneously. Each conceptual prompt can either add or remove their concept depending on if the guidance is applied positively or negatively.
Unlike Pix2Pix Zero or Attend and Excite, SEGA directly interacts with the diffusion process instead of performing any explicit gradient-based optimization.
## Self-attention Guidance (SAG)
[Paper](https://arxiv.org/abs/2210.00939)
[Self-attention Guidance](../api/pipelines/self_attention_guidance) improves the general quality of images.
SAG provides guidance from predictions not conditioned on high-frequency details to fully conditioned images. The high frequency details are extracted out of the UNet self-attention maps.
[Depth2Image](../api/pipelines/stable_diffusion/depth2img) is fine-tuned from Stable Diffusion to better preserve semantics for text guided image variation.
It conditions on a monocular depth estimate of the original image.
## MultiDiffusion Panorama
[Paper](https://arxiv.org/abs/2302.08113)
[MultiDiffusion Panorama](../api/pipelines/panorama) defines a new generation process over a pre-trained diffusion model. This process binds together multiple diffusion generation methods that can be readily applied to generate high quality and diverse images. Results adhere to user-provided controls, such as desired aspect ratio (e.g., panorama), and spatial guiding signals, ranging from tight segmentation masks to bounding boxes.
MultiDiffusion Panorama allows to generate high-quality images at arbitrary aspect ratios (e.g., panoramas).
## Fine-tuning your own models
In addition to pre-trained models, Diffusers has training scripts for fine-tuning models on user-provided data.
## DreamBooth
[Project](https://dreambooth.github.io/)
[DreamBooth](../training/dreambooth) fine-tunes a model to teach it about a new subject. I.e. a few pictures of a person can be used to generate images of that person in different styles.
## Textual Inversion
[Paper](https://arxiv.org/abs/2208.01618)
[Textual Inversion](../training/text_inversion) fine-tunes a model to teach it about a new concept. I.e. a few pictures of a style of artwork can be used to generate images in that style.
## ControlNet
[Paper](https://arxiv.org/abs/2302.05543)
[ControlNet](../api/pipelines/controlnet) is an auxiliary network which adds an extra condition.
There are 8 canonical pre-trained ControlNets trained on different conditionings such as edge detection, scribbles,
depth maps, and semantic segmentations.
## Prompt Weighting
[Prompt weighting](../using-diffusers/weighted_prompts) is a simple technique that puts more attention weight on certain parts of the text
input.
## Custom Diffusion
[Paper](https://arxiv.org/abs/2212.04488)
[Custom Diffusion](../training/custom_diffusion) only fine-tunes the cross-attention maps of a pre-trained
text-to-image diffusion model. It also allows for additionally performing Textual Inversion. It supports
multi-concept training by design. Like DreamBooth and Textual Inversion, Custom Diffusion is also used to
teach a pre-trained text-to-image diffusion model about new concepts to generate outputs involving the
concept(s) of interest.
## Model Editing
[Paper](https://arxiv.org/abs/2303.08084)
The [text-to-image model editing pipeline](../api/pipelines/model_editing) helps you mitigate some of the incorrect implicit assumptions a pre-trained text-to-image
diffusion model might make about the subjects present in the input prompt. For example, if you prompt Stable Diffusion to generate images for "A pack of roses", the roses in the generated images
are more likely to be red. This pipeline helps you change that assumption.
## DiffEdit
[Paper](https://arxiv.org/abs/2210.11427)
[DiffEdit](../api/pipelines/diffedit) allows for semantic editing of input images along with
input prompts while preserving the original input images as much as possible.
## T2I-Adapter
[Paper](https://arxiv.org/abs/2302.08453)
[T2I-Adapter](../api/pipelines/stable_diffusion/adapter) is an auxiliary network which adds an extra condition.
There are 8 canonical pre-trained adapters trained on different conditionings such as edge detection, sketch,
depth maps, and semantic segmentations.
## Fabric
[Paper](https://arxiv.org/abs/2307.10159)
[Fabric](https://github.com/huggingface/diffusers/tree/442017ccc877279bcf24fbe92f92d3d0def191b6/examples/community#stable-diffusion-fabric-pipeline) is a training-free
approach applicable to a wide range of popular diffusion models, which exploits
the self-attention layer present in the most widely used architectures to condition
the diffusion process on a set of feedback images.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# ControlNet
ControlNet is a type of model for controlling image diffusion models by conditioning the model with an additional input image. There are many types of conditioning inputs (canny edge, user sketching, human pose, depth, and more) you can use to control a diffusion model. This is hugely useful because it affords you greater control over image generation, making it easier to generate specific images without experimenting with different text prompts or denoising values as much.
<Tip>
Check out Section 3.5 of the [ControlNet](https://huggingface.co/papers/2302.05543) paper v1 for a list of ControlNet implementations on various conditioning inputs. You can find the official Stable Diffusion ControlNet conditioned models on [lllyasviel](https://huggingface.co/lllyasviel)'s Hub profile, and more [community-trained](https://huggingface.co/models?other=stable-diffusion&other=controlnet) ones on the Hub.
For Stable Diffusion XL (SDXL) ControlNet models, you can find them on the 🤗 [Diffusers](https://huggingface.co/diffusers) Hub organization, or you can browse [community-trained](https://huggingface.co/models?other=stable-diffusion-xl&other=controlnet) ones on the Hub.
</Tip>
A ControlNet model has two sets of weights (or blocks) connected by a zero-convolution layer:
- a *locked copy* keeps everything a large pretrained diffusion model has learned
- a *trainable copy* is trained on the additional conditioning input
Since the locked copy preserves the pretrained model, training and implementing a ControlNet on a new conditioning input is as fast as finetuning any other model because you aren't training the model from scratch.
This guide will show you how to use ControlNet for text-to-image, image-to-image, inpainting, and more! There are many types of ControlNet conditioning inputs to choose from, but in this guide we'll only focus on several of them. Feel free to experiment with other conditioning inputs!
Before you begin, make sure you have the following libraries installed:
```py
# uncomment to install the necessary libraries in Colab
For text-to-image, you normally pass a text prompt to the model. But with ControlNet, you can specify an additional conditioning input. Let's condition the model with a canny image, a white outline of an image on a black background. This way, the ControlNet can use the canny image as a control to guide the model to generate an image with the same outline.
Load an image and use the [opencv-python](https://github.com/opencv/opencv-python) library to extract the canny image:
Next, load a ControlNet model conditioned on canny edge detection and pass it to the [`StableDiffusionControlNetPipeline`]. Use the faster [`UniPCMultistepScheduler`] and enable model offloading to speed up inference and reduce memory usage.
For image-to-image, you'd typically pass an initial image and a prompt to the pipeline to generate a new image. With ControlNet, you can pass an additional conditioning input to guide the model. Let's condition the model with a depth map, an image which contains spatial information. This way, the ControlNet can use the depth map as a control to guide the model to generate an image that preserves spatial information.
You'll use the [`StableDiffusionControlNetImg2ImgPipeline`] for this task, which is different from the [`StableDiffusionControlNetPipeline`] because it allows you to pass an initial image as the starting point for the image generation process.
Load an image and use the `depth-estimation` [`~transformers.Pipeline`] from 🤗 Transformers to extract the depth map of an image:
Next, load a ControlNet model conditioned on depth maps and pass it to the [`StableDiffusionControlNetImg2ImgPipeline`]. Use the faster [`UniPCMultistepScheduler`] and enable model offloading to speed up inference and reduce memory usage.
For inpainting, you need an initial image, a mask image, and a prompt describing what to replace the mask with. ControlNet models allow you to add another control image to condition a model with. Let’s condition the model with an inpainting mask. This way, the ControlNet can use the inpainting mask as a control to guide the model to generate an image within the mask area.
Create a function to prepare the control image from the initial and mask images. This'll create a tensor to mark the pixels in `init_image` as masked if the corresponding pixel in `mask_image` is over a certain threshold.
Load a ControlNet model conditioned on inpainting and pass it to the [`StableDiffusionControlNetInpaintPipeline`]. Use the faster [`UniPCMultistepScheduler`] and enable model offloading to speed up inference and reduce memory usage.
[Guess mode](https://github.com/lllyasviel/ControlNet/discussions/188) does not require supplying a prompt to a ControlNet at all! This forces the ControlNet encoder to do its best to "guess" the contents of the input control map (depth map, pose estimation, canny edge, etc.).
Guess mode adjusts the scale of the output residuals from a ControlNet by a fixed ratio depending on the block depth. The shallowest `DownBlock` corresponds to 0.1, and as the blocks get deeper, the scale increases exponentially such that the scale of the `MidBlock` output becomes 1.0.
<Tip>
Guess mode does not have any impact on prompt conditioning and you can still provide a prompt if you want.
</Tip>
Set `guess_mode=True` in the pipeline, and it is [recommended](https://github.com/lllyasviel/ControlNet#guess-mode--non-prompt-mode) to set the `guidance_scale` value between 3.0 and 5.0.
<figcaptionclass="mt-2 text-center text-sm text-gray-500">guess mode without prompt</figcaption>
</div>
</div>
## ControlNet with Stable Diffusion XL
There aren't too many ControlNet models compatible with Stable Diffusion XL (SDXL) at the moment, but we've trained two full-sized ControlNet models for SDXL conditioned on canny edge detection and depth maps. We're also experimenting with creating smaller versions of these SDXL-compatible ControlNet models so it is easier to run on resource-constrained hardware. You can find these checkpoints on the [🤗 Diffusers Hub organization](https://huggingface.co/diffusers)!
Let's use a SDXL ControlNet conditioned on canny images to generate an image. Start by loading an image and prepare the canny image:
Load a SDXL ControlNet model conditioned on canny edge detection and pass it to the [`StableDiffusionXLControlNetPipeline`]. You can also enable model offloading to reduce memory usage.
Now pass your prompt (and optionally a negative prompt if you're using one) and canny image to the pipeline:
<Tip>
The [`controlnet_conditioning_scale`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/controlnet#diffusers.StableDiffusionControlNetPipeline.__call__.controlnet_conditioning_scale) parameter determines how much weight to assign to the conditioning inputs. A value of 0.5 is recommended for good generalization, but feel free to experiment with this number!
</Tip>
```py
prompt="aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
negative_prompt='low quality, bad quality, sketches'
You can use a refiner model with `StableDiffusionXLControlNetPipeline` to improve image quality, just like you can with a regular `StableDiffusionXLPipeline`.
See the [Refine image quality](./sdxl#refine-image-quality) section to learn how to use the refiner model.
Make sure to use `StableDiffusionXLControlNetPipeline` and pass `image` and `controlnet_conditioning_scale`.
```py
base=StableDiffusionXLControlNetPipeline(...)
image=base(
prompt=prompt,
controlnet_conditioning_scale=0.5,
image=canny_image,
num_inference_steps=40,
denoising_end=0.8,
output_type="latent",
).images
# rest exactly as with StableDiffusionXLPipeline
```
</Tip>
## MultiControlNet
<Tip>
Replace the SDXL model with a model like [stable-diffusion-v1-5/stable-diffusion-v1-5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) to use multiple conditioning inputs with Stable Diffusion models.
</Tip>
You can compose multiple ControlNet conditionings from different image inputs to create a *MultiControlNet*. To get better results, it is often helpful to:
1. mask conditionings such that they don't overlap (for example, mask the area of a canny image where the pose conditioning is located)
2. experiment with the [`controlnet_conditioning_scale`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/controlnet#diffusers.StableDiffusionControlNetPipeline.__call__.controlnet_conditioning_scale) parameter to determine how much weight to assign to each conditioning input
In this example, you'll combine a canny image and a human pose estimation image to generate a new image.
Load a list of ControlNet models that correspond to each conditioning, and pass them to the [`StableDiffusionXLControlNetPipeline`]. Use the faster [`UniPCMultistepScheduler`] and enable model offloading to reduce memory usage.
Diffusers' pipelines can be used as an inference engine for a server. It supports concurrent and multithreaded requests to generate images that may be requested by multiple users at the same time.
This guide will show you how to use the [`StableDiffusion3Pipeline`] in a server, but feel free to use any pipeline you want.
Start by navigating to the `examples/server` folder and installing all of the dependencies.
```py
pipinstall.
pipinstall-frequirements.txt
```
Launch the server with the following command.
```py
pythonserver.py
```
The server is accessed at http://localhost:8000. You can curl this model with the following command.
```
curl -X POST -H "Content-Type: application/json" --data '{"model": "something", "prompt": "a kitten in front of a fireplace"}' http://localhost:8000/v1/images/generations
```
If you need to upgrade some dependencies, you can use either [pip-tools](https://github.com/jazzband/pip-tools) or [uv](https://github.com/astral-sh/uv). For example, upgrade the dependencies with `uv` using the following command.
The `generate_image` function is defined as asynchronous with the [async](https://fastapi.tiangolo.com/async/) keyword so that FastAPI knows that whatever is happening in this function won't necessarily return a result right away. Once it hits some point in the function that it needs to await some other [Task](https://docs.python.org/3/library/asyncio-task.html#asyncio.Task), the main thread goes back to answering other HTTP requests. This is shown in the code below with the [await](https://fastapi.tiangolo.com/async/#async-and-await) keyword.
At this point, the execution of the pipeline function is placed onto a [new thread](https://docs.python.org/3/library/asyncio-eventloop.html#asyncio.loop.run_in_executor), and the main thread performs other things until a result is returned from the `pipeline`.
Another important aspect of this implementation is creating a `pipeline` from `shared_pipeline`. The goal behind this is to avoid loading the underlying model more than once onto the GPU while still allowing for each new request that is running on a separate thread to have its own generator and scheduler. The scheduler, in particular, is not thread-safe, and it will cause errors like: `IndexError: index 21 is out of bounds for dimension 0 with size 21` if you try to use the same scheduler across multiple threads.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Load community pipelines and components
[[open-in-colab]]
## Community pipelines
> [!TIP] Take a look at GitHub Issue [#841](https://github.com/huggingface/diffusers/issues/841) for more context about why we're adding community pipelines to help everyone easily share their work without being slowed down.
Community pipelines are any [`DiffusionPipeline`] class that are different from the original paper implementation (for example, the [`StableDiffusionControlNetPipeline`] corresponds to the [Text-to-Image Generation with ControlNet Conditioning](https://arxiv.org/abs/2302.05543) paper). They provide additional functionality or extend the original implementation of a pipeline.
There are many cool community pipelines like [Marigold Depth Estimation](https://github.com/huggingface/diffusers/tree/main/examples/community#marigold-depth-estimation) or [InstantID](https://github.com/huggingface/diffusers/tree/main/examples/community#instantid-pipeline), and you can find all the official community pipelines [here](https://github.com/huggingface/diffusers/tree/main/examples/community).
There are two types of community pipelines, those stored on the Hugging Face Hub and those stored on Diffusers GitHub repository. Hub pipelines are completely customizable (scheduler, models, pipeline code, etc.) while Diffusers GitHub pipelines are only limited to custom pipeline code.
| | GitHub community pipeline | HF Hub community pipeline |
| review process | open a Pull Request on GitHub and undergo a review process from the Diffusers team before merging; may be slower | upload directly to a Hub repository without any review; this is the fastest workflow |
| visibility | included in the official Diffusers repository and documentation | included on your HF Hub profile and relies on your own usage/promotion to gain visibility |
<hfoptionsid="community">
<hfoptionid="Hub pipelines">
To load a Hugging Face Hub community pipeline, pass the repository id of the community pipeline to the `custom_pipeline` argument and the model repository where you'd like to load the pipeline weights and components from. For example, the example below loads a dummy pipeline from [hf-internal-testing/diffusers-dummy-pipeline](https://huggingface.co/hf-internal-testing/diffusers-dummy-pipeline/blob/main/pipeline.py) and the pipeline weights and components from [google/ddpm-cifar10-32](https://huggingface.co/google/ddpm-cifar10-32):
> [!WARNING]
> By loading a community pipeline from the Hugging Face Hub, you are trusting that the code you are loading is safe. Make sure to inspect the code online before loading and running it automatically!
To load a GitHub community pipeline, pass the repository id of the community pipeline to the `custom_pipeline` argument and the model repository where you you'd like to load the pipeline weights and components from. You can also load model components directly. The example below loads the community [CLIP Guided Stable Diffusion](https://github.com/huggingface/diffusers/tree/main/examples/community#clip-guided-stable-diffusion) pipeline and the CLIP model components.
Community pipelines can also be loaded from a local file if you pass a file path instead. The path to the passed directory must contain a pipeline.py file that contains the pipeline class.
```py
pipeline=DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
custom_pipeline="./path/to/pipeline_directory/",
clip_model=clip_model,
feature_extractor=feature_extractor,
use_safetensors=True,
)
```
### Load from a specific version
By default, community pipelines are loaded from the latest stable version of Diffusers. To load a community pipeline from another version, use the `custom_revision` parameter.
<hfoptionsid="version">
<hfoptionid="main">
For example, to load from the main branch:
```py
pipeline=DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
custom_pipeline="clip_guided_stable_diffusion",
custom_revision="main",
clip_model=clip_model,
feature_extractor=feature_extractor,
use_safetensors=True,
)
```
</hfoption>
<hfoptionid="older version">
For example, to load from a previous version of Diffusers like v0.25.0:
```py
pipeline=DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
custom_pipeline="clip_guided_stable_diffusion",
custom_revision="v0.25.0",
clip_model=clip_model,
feature_extractor=feature_extractor,
use_safetensors=True,
)
```
</hfoption>
</hfoptions>
### Load with from_pipe
Community pipelines can also be loaded with the [`~DiffusionPipeline.from_pipe`] method which allows you to load and reuse multiple pipelines without any additional memory overhead (learn more in the [Reuse a pipeline](./loading#reuse-a-pipeline) guide). The memory requirement is determined by the largest single pipeline loaded.
For example, let's load a community pipeline that supports [long prompts with weighting](https://github.com/huggingface/diffusers/tree/main/examples/community#long-prompt-weighting-stable-diffusion) from a Stable Diffusion pipeline.
Community pipelines are a really fun and creative way to extend the capabilities of the original pipeline with new and unique features. You can find all community pipelines in the [diffusers/examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) folder with inference and training examples for how to use them.
This section showcases a couple of the community pipelines and hopefully it'll inspire you to create your own (feel free to open a PR for your community pipeline and ping us for a review)!
> [!TIP]
> The [`~DiffusionPipeline.from_pipe`] method is particularly useful for loading community pipelines because many of them don't have pretrained weights and add a feature on top of an existing pipeline like Stable Diffusion or Stable Diffusion XL. You can learn more about the [`~DiffusionPipeline.from_pipe`] method in the [Load with from_pipe](custom_pipeline_overview#load-with-from_pipe) section.
<hfoptionsid="community">
<hfoptionid="Marigold">
[Marigold](https://marigoldmonodepth.github.io/) is a depth estimation diffusion pipeline that uses the rich existing and inherent visual knowledge in diffusion models. It takes an input image and denoises and decodes it into a depth map. Marigold performs well even on images it hasn't seen before.
[HD-Painter](https://hf.co/papers/2312.14091) is a high-resolution inpainting pipeline. It introduces a *Prompt-Aware Introverted Attention (PAIntA)* layer to better align a prompt with the area to be inpainted, and *Reweighting Attention Score Guidance (RASG)* to keep the latents more prompt-aligned and within their trained domain to generate realistc images.
Community components allow users to build pipelines that may have customized components that are not a part of Diffusers. If your pipeline has custom components that Diffusers doesn't already support, you need to provide their implementations as Python modules. These customized components could be a VAE, UNet, and scheduler. In most cases, the text encoder is imported from the Transformers library. The pipeline code itself can also be customized.
This section shows how users should use community components to build a community pipeline.
You'll use the [showlab/show-1-base](https://huggingface.co/showlab/show-1-base) pipeline checkpoint as an example.
1. Import and load the text encoder from Transformers:
In steps 4 and 5, the custom [UNet](https://github.com/showlab/Show-1/blob/main/showone/models/unet_3d_condition.py) and [pipeline](https://huggingface.co/sayakpaul/show-1-base-with-code/blob/main/unet/showone_unet_3d_condition.py) implementation must match the format shown in their files for this example to work.
</Tip>
4. Now you'll load a [custom UNet](https://github.com/showlab/Show-1/blob/main/showone/models/unet_3d_condition.py), which in this example, has already been implemented in [showone_unet_3d_condition.py](https://huggingface.co/sayakpaul/show-1-base-with-code/blob/main/unet/showone_unet_3d_condition.py) for your convenience. You'll notice the [`UNet3DConditionModel`] class name is changed to `ShowOneUNet3DConditionModel` because [`UNet3DConditionModel`] already exists in Diffusers. Any components needed for the `ShowOneUNet3DConditionModel` class should be placed in showone_unet_3d_condition.py.
Once this is done, you can initialize the UNet:
```python
from showone_unet_3d_condition import ShowOneUNet3DConditionModel
5. Finally, you'll load the custom pipeline code. For this example, it has already been created for you in [pipeline_t2v_base_pixel.py](https://huggingface.co/sayakpaul/show-1-base-with-code/blob/main/pipeline_t2v_base_pixel.py). This script contains a custom `TextToVideoIFPipeline` class for generating videos from text. Just like the custom UNet, any code needed for the custom pipeline to work should go in pipeline_t2v_base_pixel.py.
Once everything is in place, you can initialize the `TextToVideoIFPipeline` with the `ShowOneUNet3DConditionModel`:
Push the pipeline to the Hub to share with the community!
```python
pipeline.push_to_hub("custom-t2v-pipeline")
```
After the pipeline is successfully pushed, you need to make a few changes:
1. Change the `_class_name` attribute in [model_index.json](https://huggingface.co/sayakpaul/show-1-base-with-code/blob/main/model_index.json#L2) to `"pipeline_t2v_base_pixel"` and `"TextToVideoIFPipeline"`.
2. Upload `showone_unet_3d_condition.py` to the [unet](https://huggingface.co/sayakpaul/show-1-base-with-code/blob/main/unet/showone_unet_3d_condition.py) subfolder.
3. Upload `pipeline_t2v_base_pixel.py` to the pipeline [repository](https://huggingface.co/sayakpaul/show-1-base-with-code/tree/main).
To run inference, add the `trust_remote_code` argument while initializing the pipeline to handle all the "magic" behind the scenes.
> [!WARNING]
> As an additional precaution with `trust_remote_code=True`, we strongly encourage you to pass a commit hash to the `revision` parameter in [`~DiffusionPipeline.from_pretrained`] to make sure the code hasn't been updated with some malicious new lines of code (unless you fully trust the model owners).
As an additional reference, take a look at the repository structure of [stabilityai/japanese-stable-diffusion-xl](https://huggingface.co/stabilityai/japanese-stable-diffusion-xl/) which also uses the `trust_remote_code` feature.
